Python
Make the per-reply output-token limit configurable (OPE-177)
Devika Verma2c91bd9a2235
diff --git a/coworker/agent.py b/coworker/agent.pyindex 9479718c..8ea2100e 100644--- a/coworker/agent.py+++ b/coworker/agent.py@@ -264,6 +264,11 @@ def build_engine(workspace_trusted = bool(ws and WorkspaceTrustStore().is_trusted(ws))config = load_config(ws, workspace_trusted=workspace_trusted)+ # OPE-177: the configured per-reply output ceiling rides `model_settings`, which+ # the engine spreads into every provider call (and explorer subagents inherit).+ # An explicit `max_tokens` from the caller wins over the config value.+ if config.max_output_tokens is not None and "max_tokens" not in (model_settings or {}):+ model_settings = {**(model_settings or {}), "max_tokens": config.max_output_tokens}executor = LocalExecutor(cwd=ws) if ws is not None else Nonetodo = TodoList()context = AgentContext(diff --git a/coworker/config.py b/coworker/config.pyindex 43fe33fa..60cb69e3 100644--- a/coworker/config.py+++ b/coworker/config.py@@ -9,6 +9,8 @@ workspace path. Other permission grants remain global-only.from __future__ import annotations+import os+try:import tomllib # stdlib since 3.11except ModuleNotFoundError: # 3.10, the floor requires-python declares@@ -32,6 +34,12 @@ class Config:model: str = "gpt-5.6-sol"mode: str = "interactive"max_iterations: int = 150+ # Per-reply output-token ceiling sent to the provider as `max_tokens` (thinking,+ # visible text and tool-call arguments all count against it). Unset = each+ # provider's own default (32,000 for Anthropic and OpenAI-compatible endpoints;+ # Bedrock 4,096). Anthropic recommends ~64,000 at high effort. Environment override:+ # COWORKER_MAX_OUTPUT_TOKENS. Explicit `build_engine(model_settings=...)` wins.+ max_output_tokens: Optional[int] = Noneallowed_commands: list[str] = field(default_factory=lambda: list(DEFAULT_ALLOWED_COMMANDS))@@ -79,6 +87,7 @@ _FIELDS = {"model","mode","max_iterations",+ "max_output_tokens","allowed_commands","auto_allow","allowed_domains",@@ -129,6 +138,21 @@ def workspace_allowed_commands(workspace: str | Path) -> list[str]:return list(dict.fromkeys(v.strip() for v in value if isinstance(v, str) and v.strip()))+MAX_OUTPUT_TOKENS_ENV = "COWORKER_MAX_OUTPUT_TOKENS"+++def _positive_int(value: Any, source: str) -> Optional[int]:+ """`max_output_tokens` must be a positive integer (bools are ints in Python and+ TOML `true` would otherwise pass as 1)."""+ if value is None:+ return None+ if isinstance(value, bool) or not isinstance(value, int) or value <= 0:+ raise ValueError(+ f"max_output_tokens must be a positive integer, got {value!r} ({source})"+ )+ return value++def load_config(workspace: Optional[str | Path] = None,*,@@ -154,4 +178,12 @@ def load_config([*cfg.allowed_commands, *workspace_allowed_commands(workspace)]))+ cfg.max_output_tokens = _positive_int(cfg.max_output_tokens, "config.toml")+ raw = (os.environ.get(MAX_OUTPUT_TOKENS_ENV) or "").strip()+ if raw:+ try:+ parsed: Any = int(raw)+ except ValueError:+ parsed = raw+ cfg.max_output_tokens = _positive_int(parsed, MAX_OUTPUT_TOKENS_ENV)return cfgdiff --git a/coworker/engine.py b/coworker/engine.pyindex f181c7fc..e472df8e 100644--- a/coworker/engine.py+++ b/coworker/engine.py@@ -562,6 +562,9 @@ class TurnEngine:if turn.finish_reason:# How the reply ended (OPE-173): `stop` / `tool_calls` / `length`.payload["finish_reason"] = turn.finish_reason+ if turn.output_limit:+ # The output ceiling the provider sent (OPE-177).+ payload["max_output_tokens"] = turn.output_limityield Event(EventType.ASSISTANT_MESSAGE, payload)if not turn.tool_calls:@@ -1985,10 +1988,19 @@ class TurnEngine:"""# Strip the display-only sidecars — `source` (connector cards), `_display`# (e.g. filter-hidden counts), `ts` (append-time timestamps), `reasoning`- # (thinking text), `usage` (token counts), and `finish_reason` (how the reply- # ended) — copying only messages that carry one. Whole `notice` messages- # (error/interrupted/model-switch markers) are display-only too: dropped entirely.- _SIDECARS = ("source", "_display", "ts", "reasoning", "usage", "finish_reason")+ # (thinking text), `usage` (token counts), `finish_reason` (how the reply ended)+ # and `max_output_tokens` (the ceiling sent) — copying only messages that carry+ # one. Whole `notice` messages (error/interrupted/model-switch markers) are+ # display-only too: dropped entirely.+ _SIDECARS = (+ "source",+ "_display",+ "ts",+ "reasoning",+ "usage",+ "finish_reason",+ "max_output_tokens",+ )# Auto-compaction (OPE-27): everything before the boundary is represented by the# compacted block. Outbound-only — the canonical history stays intact — and the# block+tail are byte-stable between turns, so prompt caching keeps working.@@ -2107,6 +2119,11 @@ def _assistant_message(turn: AssistantTurn, model: Optional[str] = None) -> dict# every provider call (`_outbound_messages`). The provider's raw value, where# it differs, lives in that provider's sidecar (e.g. `_anthropic.stop_reason`).message["finish_reason"] = turn.finish_reason+ if turn.output_limit:+ # The per-reply output ceiling the provider actually sent (OPE-177), so a+ # `length` finish can be read against the limit that produced it. Display+ # sidecar like `usage`: stripped before every provider call.+ message["max_output_tokens"] = turn.output_limitif turn.reasoning:# Display-only thinking text — rendered by the GUI, stripped for every provider# (`_outbound_messages`); provider-private replay blocks go via `extras` instead.diff --git a/coworker/providers/anthropic_provider.py b/coworker/providers/anthropic_provider.pyindex 0e4d59c4..7d26de17 100644--- a/coworker/providers/anthropic_provider.py+++ b/coworker/providers/anthropic_provider.py@@ -19,6 +19,8 @@ from chat.completions in ways the converters must absorb:from __future__ import annotations+import logging+import jsonimport refrom typing import Any, Optional@@ -52,6 +54,8 @@ def _usage_from(usage: Any) -> Optional[TokenUsage]:# accept ≥32k output.DEFAULT_MAX_TOKENS = 32000+logger = logging.getLogger(__name__)+# Extended thinking is ON by default (owner call 2026-07-23: no user-facing setting —# most users wouldn't know what a budget is; a per-turn composer control is future work).# The provider profile's `thinking_budget` remains a hidden override: a number replaces@@ -454,7 +458,19 @@ class AnthropicProvider(ProviderClient):# Budget must fit under max_tokens.budget = int(thinking.get("budget_tokens") or 0)floor = max(DEFAULT_MAX_TOKENS, budget + 4096)- if int(filtered.get("max_tokens") or 0) <= budget:+ requested = int(filtered.get("max_tokens") or 0)+ if requested <= budget:+ if requested:+ # A configured max_output_tokens (OPE-177) that does not clear the+ # budget would be rejected by the API; say so rather than silently+ # sending a different ceiling than the one configured.+ logger.warning(+ "max_output_tokens=%d is not above the thinking budget (%d); "+ "sending max_tokens=%d instead",+ requested,+ budget,+ floor,+ )filtered["max_tokens"] = floorif thinking.get("type") in ("enabled", "adaptive"):# Sampling knobs are rejected alongside thinking (and removed outright on 4.7+).@@ -539,6 +555,7 @@ class AnthropicProvider(ProviderClient):reasoning=_reasoning_text(thinking_blocks),extras=_anthropic_extras(thinking_blocks, stop_reason),usage=_usage_from(getattr(response, "usage", None)),+ output_limit=kwargs.get("max_tokens"),)def capabilities(self, model: str) -> ModelCapabilities:@@ -660,5 +677,6 @@ class AnthropicProvider(ProviderClient):reasoning=_reasoning_text(thinking_blocks),extras=_anthropic_extras(thinking_blocks, stop_reason),usage=usage,+ output_limit=kwargs.get("max_tokens"),))diff --git a/coworker/providers/base.py b/coworker/providers/base.pyindex 93ec660b..694b3d8c 100644--- a/coworker/providers/base.py+++ b/coworker/providers/base.py@@ -71,6 +71,12 @@ class AssistantTurn:# Token counts for this round-trip, normalized across providers. None when the# backend didn't report usage (some compat servers) — never guessed.usage: Optional[TokenUsage] = None+ # The per-reply output-token ceiling the provider actually sent on this request+ # (`max_tokens` / `max_completion_tokens` / `max_output_tokens`), after any+ # provider-side adjustment (Anthropic budget floor, OpenAI rename). None when the+ # provider left it to the server. Persisted on the assistant message as the+ # `max_output_tokens` sidecar so a run record states its ceiling (OPE-177).+ output_limit: Optional[int] = None@propertydef has_tool_calls(self) -> bool:diff --git a/coworker/providers/bedrock_provider.py b/coworker/providers/bedrock_provider.pyindex b24fb6f9..6be7c4e4 100644--- a/coworker/providers/bedrock_provider.py+++ b/coworker/providers/bedrock_provider.py@@ -63,6 +63,10 @@ def _usage_from(usage: Any) -> Optional[TokenUsage]:# Converse has no required max token param but per-model defaults vary wildly (Meta's is# 512 — an agent turn gets truncated mid-tool-call); 4096 fits every family's ceiling.DEFAULT_MAX_TOKENS = 4096+# Deliberately lower than the 32,000 of the direct Anthropic/OpenAI clients: Bedrock+# enforces per-model output ceilings that differ by family and region, and a request+# above the ceiling is a hard 400. `max_output_tokens` in config (OPE-177) raises it+# for deployments that know their models' limits.# Converse stopReason → the engine's OpenAI-shaped finish_reason vocabulary._STOP_REASON_MAP = {diff --git a/coworker/providers/openai_provider.py b/coworker/providers/openai_provider.pyindex 3e7ca306..c98d0352 100644--- a/coworker/providers/openai_provider.py+++ b/coworker/providers/openai_provider.py@@ -120,6 +120,13 @@ def _param_fix_retry(kwargs: dict[str, Any], exc: Exception) -> dict[str, Any]:raise exc+def _output_limit(kwargs: dict[str, Any]) -> Optional[int]:+ """The completion ceiling actually sent, after `_param_fix_retry` may have renamed+ `max_tokens` to `max_completion_tokens` or dropped it (server default -> None)."""+ value = kwargs.get("max_tokens", kwargs.get("max_completion_tokens"))+ return int(value) if value else None++def _usage_from(usage: Any) -> Optional[TokenUsage]:"""chat.completions usage → normalized counts. `prompt_tokens` INCLUDES cachedtokens, so the cached share is subtracted into `cache_read`; no write-side split@@ -219,6 +226,7 @@ class OpenAIProvider(ProviderClient):raw=response,reasoning=_delta_reasoning(message),usage=_usage_from(getattr(response, "usage", None)),+ output_limit=_output_limit(kwargs),)def capabilities(self, model: str) -> ModelCapabilities:@@ -317,6 +325,7 @@ class OpenAIProvider(ProviderClient):finish_reason=finish_reason,reasoning="".join(reasoning_parts) or None,usage=usage,+ output_limit=_output_limit(kwargs),))diff --git a/docs/config.example.toml b/docs/config.example.tomlindex 9638ad23..b3b6b0c9 100644--- a/docs/config.example.toml+++ b/docs/config.example.toml@@ -7,6 +7,9 @@model = "gpt-5.5" # default model id (override per session in the UI/CLI)mode = "interactive" # plan | interactive | auto | custommax_iterations = 12 # max model<->tool iterations per turn before stopping+# max_output_tokens = 64000 # per-reply output cap sent as max_tokens (thinking + text ++ # tool arguments). Unset = provider default (32000; Bedrock+ # 4096). Env override: COWORKER_MAX_OUTPUT_TOKENS.# Commands auto-allowed without an approval prompt (prefix match).allowed_commands = [diff --git a/tests/test_max_output_tokens.py b/tests/test_max_output_tokens.pynew file mode 100644index 00000000..023f8504--- /dev/null+++ b/tests/test_max_output_tokens.py@@ -0,0 +1,335 @@+"""OPE-177: the per-reply output-token limit is configurable.++`max_output_tokens` (config.toml, `COWORKER_MAX_OUTPUT_TOKENS`, or an explicit+`model_settings["max_tokens"]`) reaches every provider request as `max_tokens`; unset keeps+each provider's own default byte-for-byte. The value the provider actually sent (after the+Anthropic budget-thinking floor, or the OpenAI `max_completion_tokens` rename) comes back on+`AssistantTurn.output_limit` and is persisted on the assistant message as the+`max_output_tokens` sidecar, next to `usage` and `finish_reason`.+"""++from __future__ import annotations++import asyncio+import logging+from contextlib import contextmanager+from types import SimpleNamespace++import aisuite as ai+import pytest+from coworker.agent import build_code_engine+from coworker.config import load_config+from coworker.engine import TurnEngine, _assistant_message+from coworker.events import EventType+from coworker.permissions import PermissionEngine+from coworker.providers import (+ AnthropicProvider,+ AssistantTurn,+ ModelCapabilities,+ OpenAIProvider,+ ProviderClient,+)+from coworker.providers.anthropic_provider import DEFAULT_MAX_TOKENS as ANTHROPIC_DEFAULT+from coworker.providers.openai_provider import _MAX_TOKENS_ERROR+from coworker.providers.openai_provider import DEFAULT_MAX_TOKENS as OPENAI_DEFAULT+from coworker.tools import ToolRegistry++# -- config -------------------------------------------------------------------------+++def test_config_default_is_unset(tmp_path):+ cfg = load_config(global_path=tmp_path / "nope.toml")+ assert cfg.max_output_tokens is None+++def test_config_global_then_workspace_override(tmp_path):+ g = tmp_path / "global.toml"+ g.write_text("max_output_tokens = 64000\n")+ assert load_config(global_path=g).max_output_tokens == 64000+ ws = tmp_path / "ws"+ (ws / ".coworker").mkdir(parents=True)+ (ws / ".coworker" / "config.toml").write_text("max_output_tokens = 48000\n")+ assert load_config(ws, global_path=g).max_output_tokens == 48000+++def test_config_env_override_beats_files(tmp_path, monkeypatch):+ g = tmp_path / "global.toml"+ g.write_text("max_output_tokens = 64000\n")+ monkeypatch.setenv("COWORKER_MAX_OUTPUT_TOKENS", "70000")+ assert load_config(global_path=g).max_output_tokens == 70000+++ "toml_value,env_value",+ [("0", "0"), ("-5", "-5"), ('"lots"', "lots"), ("1.5", "1.5"), ("true", "true")],+)+def test_config_rejects_non_positive_or_non_integer(tmp_path, monkeypatch, toml_value, env_value):+ g = tmp_path / "global.toml"+ g.write_text(f"max_output_tokens = {toml_value}\n")+ with pytest.raises(ValueError, match="max_output_tokens"):+ load_config(global_path=g)+ g.write_text("")+ monkeypatch.setenv("COWORKER_MAX_OUTPUT_TOKENS", env_value)+ with pytest.raises(ValueError, match="max_output_tokens"):+ load_config(global_path=g)+++# -- build_engine -------------------------------------------------------------------+++class _Stub(ProviderClient):+ def complete(self, **k): # pragma: no cover+ raise NotImplementedError++ def capabilities(self, m):+ return ModelCapabilities()+++def _engine_with_config(tmp_path, toml: str, **kwargs):+ (tmp_path / ".coworker").mkdir(exist_ok=True)+ (tmp_path / ".coworker" / "config.toml").write_text(toml)+ return build_code_engine(workspace=tmp_path, provider=_Stub(), **kwargs)+++def test_build_engine_passes_config_value_as_max_tokens(tmp_path):+ engine = _engine_with_config(tmp_path, "max_output_tokens = 64000\n")+ try:+ assert engine.model_settings["max_tokens"] == 64000+ finally:+ engine.executor.close()+++def test_build_engine_explicit_model_settings_win_over_config(tmp_path):+ engine = _engine_with_config(+ tmp_path, "max_output_tokens = 64000\n", model_settings={"max_tokens": 1000}+ )+ try:+ assert engine.model_settings["max_tokens"] == 1000+ finally:+ engine.executor.close()+++def test_build_engine_unset_leaves_model_settings_empty(tmp_path):+ engine = _engine_with_config(tmp_path, "max_iterations = 3\n")+ try:+ assert engine.model_settings == {}+ finally:+ engine.executor.close()+++# -- Anthropic ----------------------------------------------------------------------+++class _FakeAnthropic:+ def __init__(self, response=None, events=None):+ self.kwargs: dict = {}++ def create(**kwargs):+ self.kwargs = kwargs+ if kwargs.get("stream"):+ return iter(events or [])+ return response++ @contextmanager+ def stream(**kwargs):+ self.kwargs = kwargs+ yield SimpleNamespace(get_final_message=lambda: response)++ self.messages = SimpleNamespace(create=create, stream=stream)+ self.beta = SimpleNamespace(messages=SimpleNamespace(create=create, stream=stream))+++def _anthropic_response(stop_reason="end_turn"):+ return SimpleNamespace(+ content=[SimpleNamespace(type="text", text="hi")], stop_reason=stop_reason+ )+++def _anthropic_events():+ return [+ SimpleNamespace(+ type="content_block_start",+ index=0,+ content_block=SimpleNamespace(type="text", text=""),+ ),+ SimpleNamespace(+ type="content_block_delta",+ index=0,+ delta=SimpleNamespace(type="text_delta", text="hi"),+ ),+ SimpleNamespace(type="message_delta", delta=SimpleNamespace(stop_reason="end_turn")),+ ]+++def test_anthropic_complete_and_stream_send_configured_limit():+ fake = _FakeAnthropic(response=_anthropic_response(), events=_anthropic_events())+ provider = AnthropicProvider(client=fake)+ turn = provider.complete(+ model="m", messages=[{"role": "user", "content": "x"}], max_tokens=64000+ )+ assert fake.kwargs["max_tokens"] == 64000+ assert turn.output_limit == 64000++ chunks = list(+ provider.stream(+ model="m", messages=[{"role": "user", "content": "x"}], max_tokens=64000+ )+ )+ assert fake.kwargs["max_tokens"] == 64000+ assert chunks[-1].turn.output_limit == 64000+++def test_anthropic_default_path_unchanged():+ fake = _FakeAnthropic(response=_anthropic_response())+ turn = AnthropicProvider(client=fake).complete(+ model="m", messages=[{"role": "user", "content": "x"}]+ )+ assert fake.kwargs["max_tokens"] == ANTHROPIC_DEFAULT+ assert turn.output_limit == ANTHROPIC_DEFAULT+++def test_anthropic_budget_floor_overrides_small_value_and_warns(caplog):+ fake = _FakeAnthropic(response=_anthropic_response())+ provider = AnthropicProvider(client=fake, thinking_budget=8192)+ with caplog.at_level(logging.WARNING, logger="coworker.providers.anthropic_provider"):+ turn = provider.complete(+ model="claude-haiku-4-5",+ messages=[{"role": "user", "content": "x"}],+ max_tokens=4000,+ )+ floor = max(ANTHROPIC_DEFAULT, 8192 + 4096)+ assert fake.kwargs["max_tokens"] == floor+ assert turn.output_limit == floor+ assert any("max_output_tokens" in r.getMessage() for r in caplog.records)++ caplog.clear()+ with caplog.at_level(logging.WARNING, logger="coworker.providers.anthropic_provider"):+ turn = provider.complete(+ model="claude-haiku-4-5",+ messages=[{"role": "user", "content": "x"}],+ max_tokens=64000,+ )+ assert fake.kwargs["max_tokens"] == 64000+ assert turn.output_limit == 64000+ assert not caplog.records+++# -- OpenAI-compatible --------------------------------------------------------------+++class _FakeOpenAI:+ """Records every create() kwargs; can reject `max_tokens` once like a reasoning model."""++ def __init__(self, *, stream_chunks=None, response=None, reject_max_tokens_once=False):+ self.calls: list[dict] = []+ self._reject = reject_max_tokens_once++ def create(**kwargs):+ self.calls.append(kwargs)+ if self._reject and "max_tokens" in kwargs:+ self._reject = False+ raise RuntimeError(+ "Unsupported parameter: " + _MAX_TOKENS_ERROR + " with this model."+ )+ return stream_chunks if kwargs.get("stream") else response++ self.chat = SimpleNamespace(completions=SimpleNamespace(create=create))+++def _openai_chunks():+ return [+ SimpleNamespace(+ choices=[+ SimpleNamespace(+ delta=SimpleNamespace(content="hi", tool_calls=None),+ finish_reason="stop",+ )+ ],+ usage=None,+ )+ ]+++def _openai_response():+ return SimpleNamespace(+ choices=[+ SimpleNamespace(+ message=SimpleNamespace(content="hi", tool_calls=None), finish_reason="stop"+ )+ ],+ usage=None,+ )+++def test_openai_stream_and_complete_send_configured_limit():+ fake = _FakeOpenAI(stream_chunks=_openai_chunks(), response=_openai_response())+ provider = OpenAIProvider(client=fake)+ chunks = list(+ provider.stream(+ model="m", messages=[{"role": "user", "content": "x"}], max_tokens=64000+ )+ )+ assert fake.calls[-1]["max_tokens"] == 64000+ assert chunks[-1].turn.output_limit == 64000++ turn = provider.complete(+ model="m", messages=[{"role": "user", "content": "x"}], max_tokens=64000+ )+ assert fake.calls[-1]["max_tokens"] == 64000+ assert turn.output_limit == 64000+++def test_openai_default_path_unchanged():+ fake = _FakeOpenAI(response=_openai_response())+ turn = OpenAIProvider(client=fake).complete(+ model="m", messages=[{"role": "user", "content": "x"}]+ )+ assert fake.calls[-1]["max_tokens"] == OPENAI_DEFAULT+ assert turn.output_limit == OPENAI_DEFAULT+++def test_openai_rename_retry_keeps_the_configured_limit():+ fake = _FakeOpenAI(response=_openai_response(), reject_max_tokens_once=True)+ turn = OpenAIProvider(client=fake).complete(+ model="m", messages=[{"role": "user", "content": "x"}], max_tokens=64000+ )+ assert len(fake.calls) == 2+ assert "max_tokens" not in fake.calls[-1]+ assert fake.calls[-1]["max_completion_tokens"] == 64000+ assert turn.output_limit == 64000+++# -- engine record ------------------------------------------------------------------+++class _LimitProvider(ProviderClient):+ def complete(self, *, model, messages, tools=None, **settings):+ return AssistantTurn(text="done", finish_reason="stop", output_limit=64000)++ def capabilities(self, model):+ return ModelCapabilities()+++def test_engine_records_output_limit_on_message_and_event(tmp_path):+ registry = ToolRegistry()+ registry.register_all(ai.toolkits.files(root=str(tmp_path), allow_write=True))+ engine = TurnEngine(+ provider=_LimitProvider(),+ registry=registry,+ permissions=PermissionEngine(workspace_root=tmp_path),+ model="gpt-5.5",+ )++ async def _collect():+ return [ev async for ev in engine.run("hello")]++ events = asyncio.run(_collect())+ assistant = next(ev for ev in events if ev.type == EventType.ASSISTANT_MESSAGE)+ assert assistant.data["max_output_tokens"] == 64000+ persisted = next(m for m in engine.messages if m.get("role") == "assistant")+ assert persisted["max_output_tokens"] == 64000+ assert all("max_output_tokens" not in m for m in engine._outbound_messages())+++def test_assistant_message_omits_limit_when_unknown():+ assert "max_output_tokens" not in _assistant_message(AssistantTurn(text="x"))